Crop diseases pose a big threat to food production around the world. Early detection of plant diseases is key to effective interventions, ultimately reducing the impact on our food supply. Deep learning excels at image classification, making it a critical tool for predicting plant diseases. This article deliberates on the role played by deep learning techniques for disease prediction in plants.
Image Credit: Tunatura/Shutterstock
Importance of Plant Disease Prediction
Plant illnesses have remained one of the major concerns in agriculture, as they reduce crop yield and quality. Plant diseases adversely impact the agricultural economy of developing countries that depend on a single or a few crops, as severe damage to entire areas of planted crops leads to huge financial losses.
For instance, diseases caused by either fungus or bacteria, such as pink rot, black scurf, silver scurf, late blight, black dot, fusarium dry rot, early blight, and common scab, significantly affect the productivity of tomatoes, potatoes, and rice. When the plant is infected by any disease, the leaves start showing different symptoms, like changes in their texture, shape, and color.
Gradually, the disease starts affecting other plant parts, which eventually impacts productivity. The diseases can be identified in the early stages and necessary actions can be implemented to prevent further loss when the plants are monitored carefully at regular intervals. This necessitated the development of reliable and available methods for disease prediction/early detection and diagnosis of diseases to enhance yield.
Currently, complex backgrounds with different real-world conditions, computational time complexity, lack of photographs labeled and collected from real-life scenarios in existing available datasets, infection level of disease and life cycle of the plant-based on infection, segmentation sensitivity towards the region of interest, and detection of multi-infection in the full single or multiple leaves sample are the major challenges in the field of plant disease prediction.
Role of Deep Learning
Deep learning techniques can be utilized for effectively predicting plant diseases as they detect the presence of diseases and pests in farms with high accuracy by overcoming the existing challenges. The need for huge datasets is the major problem with the use of deep learning. However, the advent of digital technology has eased the collection of required data and facilitated optimal decision-making for identifying different diseases in plants using deep learning techniques.
For instance, in a study published in the 2021 5th International Conference on Computing Methodologies and Communication (ICCMC), a comparative analysis of deep learning models, including Inception v3, ResNet50, and DenseNet, with and without augmentation, was performed for classifying 10 diseases of six crops.
Plant disease prediction in the early stages does not require background removal steps, infected area segmentation, and pre-processing. A public dataset containing 10,000 images of healthy and diseased plants was used for classification. The ResNet50 attained the highest accuracy among all models, achieving 98.2% and 97.3% accuracies with and without augmentation, respectively.
Applications of Deep Learning
A deep learning-based method using LeNet architecture as a convolutional neural network (CNN) has been proposed to automate the banana leaf disease classification process. Preliminary results showed the effectiveness of the approach even under challenging conditions like complex backgrounds, illumination, different resolutions, poses, sizes, and orientations of real-scene images.
A CNN-based method has been developed to identify unwanted weeds, like unwanted grasses and broadleaf weeds, in soybean fields. Drones have been used to capture images, and the database used for analysis includes fifteen thousand pictures of broadleaf, grass weeds, soybeans, soil, and other weeds.
The SafeNet architecture was used for neural network training, and the cafe software included AlexNet within it. A robust image database was built using the Pynovisao algorithm. Leveraging the CNN-based approach, an impressive 99% accuracy was achieved.
A deep learning-based approach for herbarium species identification was proposed in a study. The study primarily focused on the abilities of CNNs in the automatic identification of plant species. ImageNet classification was used as it performs efficiently in the CNN process, while transfer learning was employed for domain-related training.
The neural network achieved higher accuracy when trained on one set of species and tested on a different set. It was also shown that using a herbarium dataset, transfer learning can be feasible to another region even when the species don't match. Moreover, the pre-processing technique removed the handwritten tags and noise effectively.
Another study introduced a deep learning-based technique for image classification. This study focused on identifying individual lesions and spots instead of considering the entire leaf. Data augmentation techniques addressed the issue of a lack of a plant image database.
The accuracy was 12% higher while using only lesions and spots instead of the entire leaf. GoogleNet CNN was utilized in the experimental setup. Additionally, three types of images were used, including images with the background removed, images without any modification, and an expanded dataset.
Multiple CNN models were developed to perform plant disease diagnosis and detection using simple leaf images of diseased and healthy plants. An open database of 87,848 images consisting of 25 plants in a set of 58 distinct classes of plant and disease combinations, including healthy plants, was used for training the models.
Many model architectures were trained, and the architecture that showed the best performance attained a 99.53% success rate in identifying the corresponding plant and disease combination/healthy plant. The substantially high success rate indicates the feasibility of the model as an early warning tool. Moreover, this approach could be expanded to develop an integrated plant disease identification system that can operate in real cultivation conditions.
Although northern leaf blight (NLB) causes severe yield loss in maize, scouting large areas to diagnose the disease accurately is difficult and time-consuming. A CNN-based approach has been developed that can automatically identify NLB lesions with high reliability in field-acquired images of maize plants.
This approach utilizes a computational pipeline of CNNs that mitigates the challenges of limited data and the myriad irregularities in field-grown plant images. Many CNNs were trained to classify small regions of images as containing NLB lesions or not, and their predictions were combined into separate heat maps.
These heat maps were then fed into a final CNN trained to classify the entire image as diseased plants or healthy plants. The proposed system realized 96.7% accuracy on test set images that were not used in training. Mounting these systems on ground- or aerial-based vehicles can also assist in precision breeding for disease resistance.
Recent Development
In Saudi Arabia, the agricultural industry suffers from the impacts of vegetable diseases in the Central Province. One parasitic disease, two physiological diseases, and 32 fungal diseases primarily affect the productivity of vegetables.
In a recent study published in the International Journal of Advanced Computer Science and Applications (IJACSA), researchers developed a framework for agriculture plant disease prediction using deep learning classifiers, specifically for the agricultural industry in Saudi Arabia.
Onion, pepper, and tomatoes were selected for the study to realize early diagnosis of their diseases to boost quality and productivity. In the first stage, common image processing methods using machine learning classifiers were used. Then, hyperparameter-tuned machine learning classifiers, such as random forest, support vector machine, and k-nearest neighbor, were performed to determine an outcome.
Eventually, the proposed Deep Learning Plant Disease Detection System makes use of tuned machine learning models. In the second stage, potential CNN designs were assessed using the stochastic gradient descent and the supplied input dataset. The best CNN model was fine-tuned using many optimizers to enhance the classification accuracy. Results showed that a modified CNN attained 99.5% classification accuracy and a 1.00 F1 score for pepper disease in the first phase module. An enhanced GoogleNet using the Adam optimizer realized 99.5% classification accuracy and 0.997 F1 score for pepper illnesses, which was much higher than previous models.
Overall, deep learning offers a powerful tool for plant disease prediction with high accuracy. This technology holds great promise for improving agricultural yields and food security. However, deep learning for plant disease prediction faces hurdles like limited labeled data and complex variations in real-world field images.
References and Further Reading
Lijo, J. (2021). Analysis of effectiveness of augmentation in plant disease prediction using deep learning. 2021 5th International Conference on Computing Methodologies and Communication (ICCMC), 1654-1659. https://doi.org/10.1109/ICCMC51019.2021.9418266
Baljon, M. (2023). A Framework for Agriculture Plant Disease Prediction using Deep Learning Classifier. International Journal of Advanced Computer Science and Applications, 14(8). https://doi.org/10.14569/IJACSA.2023.01408119
Malathi, L., Yogashree, P., Thamaraiselvi, A. (2020). A Survey on Plant Disease Prediction Using Deep Learning. International Journal of Multidisciplinary Educational Research, 9(4), 101. https://www.researchgate.net/profile/J-Roopavathy/publication/351820770
G, Ushadevi., BV, Gokulnath. (2020). A survey on plant disease prediction using machine learning and deep learning techniques. Inteligencia Artificial, 23(65), 136-154. https://doi.org/10.4114/intartif.vol23iss65pp136-154